Occluded Face Recovery Using Generative Adversarial Network (GAN)

In unconstrained situations faces are easily covered/occluded by some other object, such as facial accessories. However, reconstructing that corrupted face is a very challenging task. Generative Adversarial Network (GAN) is effectively able to regain the occluded face and recover the blur and noisy images. The GAN method used one generative model and one discriminative model. In this work we used two discriminative models: first one a local discriminative model and the second one a global discriminative model. However, the generative model is used to recover the occluded image. The local and global discriminative models identify the differences between the real face image and the generative face image at the local and global level on the face. This work is implemented on the different datasets. The occlusion was created on these datasets by using two different methods: one by imposing rectangular blocks and the other by using inpainting. We use Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM), two widely used metrics, to quantitatively assess visual quality in terms of both appearance and structure. This work also shows the recognition accuracy with existing state-of-the-art methods.